面向 AI agent skills、MCP servers、prompts 和 agent-native developer tools 的每日 GitHub 雷达。
大多数开发者不缺另一个 awesome list，真正缺的是一套可重复的方法：发现需求正在哪里形成，判断什么值得做，并在开新仓库之前先发布证据。
Agent Skill Radar 把公开 GitHub 信号变成每日 build brief。中文用户是这个项目优先服务的人群：开发者、技术创作者、AI 工具训练营、独立开发者和小团队，都可以用它做选题、选型、内容和产品化验证。
| 步骤 | 做什么 | 输出 |
|---|---|---|
| 扫描 | 搜索 agent skills、MCP servers、prompts、context tools、agent-native CLIs。 | 原始仓库和 issue 数据 |
| 评分 | 按热度、活跃度、issue 需求、扩展性、创作者适配度、新颖度和饱和度排序。 | 机会分数和阶段 |
| 简报 | 把高分仓库转成可执行的 companion-tool 角度。 | 带 issue 证据的 build brief |
| 发布 | 每天提交一份 Markdown 报告。 | GitHub、博客、图文和短视频都能引用的报告 |
| 构建 | 用重复出现的强信号决定下一个小工具。 | 更少随机开坑 |
当前公开报告：reports/latest.md
首批信号示例：
| 排名 | 机会 | 构建角度 |
|---|---|---|
| 1 | addyosmani/agent-skills |
跨 agent 的 skill index、installer 或 quality benchmark |
| 2 | getsentry/XcodeBuildMCP |
MCP registry、security checker、config generator 或 compatibility layer |
| 3 | HKUDS/CLI-Anything |
带 JSON 输出和确定性工作流的 agent-native CLI wrapper |
- 想围绕 Codex、Claude Code、Copilot、Cursor、Gemini CLI 或 MCP 做工具的开发者。
- 想先看证据再决定开源项目方向的独立开发者。
- 想把 GitHub 研究变成公众号、小红书、视频、社群选题的技术创作者。
- 想筛选 MCP/Agent 工具、做内部培训或产品化验证的小团队。
git clone https://github.com/soarsky1991/skill-radar.git
cd skill-radar
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
agent-skill-radar run --limit-per-query 8 --issue-limit 4提高 GitHub API 限额：
export GITHUB_TOKEN=github_token_here输出：
data/YYYY-MM-DD.json
data/latest.json
reports/YYYY-MM-DD.md
reports/latest.md
采集评分 JSON：
agent-skill-radar collect \
--limit-per-query 12 \
--issue-limit 6 \
--days 45 \
--out data/latest.json从 JSON 渲染报告：
agent-skill-radar report \
--input data/latest.json \
--out reports/latest.md运行每日流水线：
agent-skill-radar run添加自定义 GitHub 搜索：
agent-skill-radar collect \
--query "agent memory coding assistant stars:>100 created:>=2025-01-01" \
--out data/custom.json这个分数是实用的构建信号，不是投资指标。
- Heat：stars、forks 和已有受众。
- Freshness：近期仓库活跃度。
- Velocity proxy：按创建时间估算的 stars/day。
- Issue demand：comments、reactions 和需求型 issue 标题。
- Extensibility：是否自然适合 plugin、skill、MCP、CLI、prompt 或 template。
- Creator fit：是否适合做教程、benchmark、公开报告或内容系列。
- Novelty：生态是否足够新，还有明显空位。
- Saturation penalty：生态是否已经太成熟、太拥挤。
阶段：
build-now：可以立刻做 proof of concept。probe-this-week：先发 fake-door README、issue reply 或 demo post。content-first：先做内容观察反馈，再决定要不要构建。archive：保留为参考。
- 每日：提交一份 radar report。
- 每周两次：把一个高分 gap 拆成公开内容。
- 每周：发布趋势笔记，请社区补充遗漏仓库。
- 每两周：判断最强重复信号是否值得孵化成 companion project。
- Star delta snapshots。
- Repo allowlist / denylist。
- Issue demand 聚类。
- Codex、Claude Code、Copilot、Cursor、Gemini CLI、MCP 兼容性字段。
- GitHub Pages dashboard。
- Companion repo 验证模板。
欢迎开 issue 提供：
- 应该被追踪的 repo 或生态；
- 需要过滤的噪音结果；
- 更有用的评分信号；
- 能证明真实需求的 issue、discussion 或 release 证据。
小而具体、有证据的建议最有价值。
Daily GitHub radar for AI agent skills, MCP servers, prompts, and agent-native developer tools.
Most builders do not need another giant awesome list. They need a repeatable way to notice where developer demand is forming, decide what is worth building, and publish evidence before opening another repo.
Agent Skill Radar turns public GitHub signals into a daily build brief. Chinese-speaking builders are the first audience for monetization and community growth, while the English version stays complete so the project can connect with the global open-source and AI agent ecosystem.
| Step | What happens | Output |
|---|---|---|
| Scan | Search GitHub for agent skills, MCP servers, prompts, context tools, and agent-native CLIs. | Raw repo and issue data |
| Score | Rank by heat, freshness, issue demand, extensibility, creator fit, novelty, and saturation. | Opportunity score and stage |
| Brief | Convert the top repos into concrete companion-tool angles. | Build briefs with issue evidence |
| Publish | Commit a dated Markdown report every day. | Reports for GitHub, blogs, newsletters, and short-form content |
| Build | Use repeated strong signals to choose the next small tool. | Less random shipping |
The current public report lives at reports/latest.md.
Example signal from the first run:
| Rank | Opportunity | Build angle |
|---|---|---|
| 1 | addyosmani/agent-skills |
Cross-agent skill index, installer, or quality benchmark |
| 2 | getsentry/XcodeBuildMCP |
MCP registry, security checker, config generator, or compatibility layer |
| 3 | HKUDS/CLI-Anything |
Agent-native CLI wrapper with JSON output and deterministic workflows |
- Developers building tools around Codex, Claude Code, Copilot, Cursor, Gemini CLI, or MCP.
- Indie hackers who want evidence before committing to a new open-source idea.
- Technical creators turning GitHub research into articles, videos, newsletters, or community discussions.
- Small teams evaluating MCP tools, agent workflows, and productization paths.
git clone https://github.com/soarsky1991/skill-radar.git
cd skill-radar
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
agent-skill-radar run --limit-per-query 8 --issue-limit 4Set a token for higher GitHub API limits:
export GITHUB_TOKEN=github_token_hereOutputs:
data/YYYY-MM-DD.json
data/latest.json
reports/YYYY-MM-DD.md
reports/latest.md
Collect scored JSON:
agent-skill-radar collect \
--limit-per-query 12 \
--issue-limit 6 \
--days 45 \
--out data/latest.jsonRender a report:
agent-skill-radar report \
--input data/latest.json \
--out reports/latest.mdRun the daily pipeline:
agent-skill-radar runAdd a custom GitHub search query:
agent-skill-radar collect \
--query "agent memory coding assistant stars:>100 created:>=2025-01-01" \
--out data/custom.jsonThe score is a practical build signal, not an investment metric.
- Heat: stars, forks, and existing audience.
- Freshness: recent repository activity.
- Velocity proxy: stars per day since creation.
- Issue demand: comments, reactions, and demand-shaped titles.
- Extensibility: whether the repo naturally supports plugins, skills, MCP, CLI, prompts, or templates.
- Creator fit: whether it can become a visible workflow, tutorial, benchmark, or content series.
- Novelty: whether the ecosystem is young enough to have gaps.
- Saturation penalty: whether a giant ecosystem is already too mature.
Stages:
build-now: create a proof of concept immediately.probe-this-week: publish a fake-door README, issue reply, or demo post.content-first: cover the topic, watch responses, then build.archive: keep for reference.
- Daily: commit one radar report.
- Twice weekly: turn one high-score gap into a public breakdown.
- Weekly: publish a trend note and ask for missing repos.
- Every two weeks: decide whether the strongest repeated signal deserves a companion project.
- Star delta snapshots.
- Repo allowlist and denylist.
- Issue clustering by demand type.
- Compatibility fields for Codex, Claude Code, Copilot, Cursor, Gemini CLI, and MCP.
- GitHub Pages dashboard.
- One-click brief template for validating a companion repo.
Open an issue with:
- a repo or ecosystem you think should be tracked;
- a noisy result that should be filtered;
- a scoring signal that would make the radar more useful;
- issue, discussion, or release evidence that shows real demand.
Small, evidence-backed suggestions are more useful than broad category requests.
MIT